Machine learning-driven surrogate model development for geotechnical numerical simulation

Machine learning techniques establish the relationship between inputs and outputs in numerical simulations, circumventing the need for complex modelling and post-processing. This reduces the expertise and time required, facilitating wider adoption of numerical simulation methods. In this paper, taking the classic geotechnical engineering problem of slope safety factor calculation as an example, a comprehensive methodology for optimising numerical simulations using machine learning is presented. This includes: (a) the determination and quantification of input and output parameters for numerical simulations; (b) the design of the neural network, including the algorithm selection, the structure design, and the activation functions selection; (c) the design of training standards for neural networks; (d) the design of training and test sets using orthogonal and full factorial design methods; and (e) the model performance evaluation and the characteristics of prediction errors analysis. Furthermore, the probability of achieving the acceptable models on a training set and the extrapolation performance of the surrogate model are discussed in detail, which are beneficial for the design of the training sets and training sessions. The paper aims to standardise building and evaluating geotechnical numerical simulation surrogate models using machine learning, easing their application in engineering practice.